Authors: Henry Han, Luyu Niu, Diane Li, Jiacun Wang
Organizations: Data Science and AI Innovation Laboratory, Department of Computer Science, Baylor University, Waco, TX, 76798, USA · Department of Business, Management and Accounting, University of Maryland Eastern Shore, Princess Anne, MD, 21853, USA · Department of Computer Science and Software Engineering, Monmouth University, West Long Branch, NJ, 07764, USA
Manifold learning methods like t-SNE and UMAP preserve local neighborhoods, but they do not preserve the ranking of the neighbors inside them. Two embeddings can return the same neighbor set while disagreeing on which point is closest. This loss of order degrades downstream applications that value the order in the embedding. We propose Smooth Rank t-SNE (SR-tSNE), which adds a differentiable rank term to the standard t-SNE loss. To enable efficient training, this term uses a sigmoid rank approximation to explicitly force a point's true high-dimensional neighbors to outrank unrelated intruders and retain their original order. We also introduce Ordinal DLP (O-DLP), an extension of the Degree of Locality Preservation (DLP) metric that credits a neighbor only when it returns at its original rank. Across eight benchmarks spanning image, text, speech, remote-sensing and single-cell data, SR-tSNE outperforms 12 baselines in rank preservation and neighborhood overlap. Ablation studies demonstrate our rank term alone boosts O-DLP by 5% to 28%, a gain not achievable using standard pairwise loss. This advantage transfers to downstream applications. SR-tSNE achieves the best label-transfer performance across seven vision datasets. For cell-type discovery across six single-cell datasets, it ranks first in 17 of 18 dataset-metric combinations, whether clustering resolution is selected with or without cell-type labels. In intraday trading, SR-tSNE achieves higher returns, lower risk, and lower turnover than the baselines. To our knowledge, SR-tSNE is the first t-SNE-family method with explicit order preservation.
Figures & tables
Figure 1: Mechanism of the smooth rank regularization on three MNIST queries ( k=6 ). Solid discs show true input-space neighbors positioned in the embedding space and labeled by original rank: teal preserves exact rank, while coral is displaced. Dashed circles mark embedding-space intruders (from outside the top- 30 ). Lrank simultaneously repels intruders and restores neighbor order, improving exact-rank preservation (O-DLP) by 19% over t-SNE on the full 70,000 -sample MNIST dataset (Table 4 ); the gain over the equal-budget control is +17% (Table 2 ).
Dataset
Domain
n
p
Classes
MNIST
Vision
70 000
784
10
CIFAR-10
Vision
60 000
512
10
Fashion-MNIST
Vision
70 000
784
10
Speech Commands
Speech
59 387
768
35
Renal-Cortex
Single-cell
50 027
36 030
35
Brain-Cortex
Single-cell
50 212
58 232
35
Table 1: Overview of the eight benchmark datasets. Sources and data types are listed in Table A.11 of A.2 .
Dataset
t-SNE
DREAMS
λ=0 control
SR-tSNE
Gain (%)
MNIST
.0781
.0409
.0793
.0927
+17
CIFAR-10
.0545
.0254
.0548
.0597
+9
Fashion-MNIST
.0672
.0497
.0672
.0792
+18
Speech Commands
.0809
.0533
.0818
.0858
+5
Renal-Cortex
.0578
.0336
.0578
.0741
+28
Brain-Cortex
.0615
.0241
.0616
.0767
+25
Table 2: O-DLP at k=10 for t-SNE, the strongest published baseline (DREAMS), the equal-budget λ=0 control and SR-tSNE, with the mean over datasets and the gain of the rank term over the control. All twelve methods are in Table 4 below.
Figure 2: O-DLP versus neighborhood size k on the eight benchmarks of Table 1 . Each panel carries every method run on that dataset, so the curves match the entries of Appendix Table 4 .
Method
MN
CF
FM
SC
RC
BC
CT
AG
Mean
TriMap
1.04
1.00
1.06
1.16
1.07
1.05
1.10
1.09
1.07
PaCMAP
1.10
1.02
1.08
1.13
1.07
1.05
1.25
1.08
1.10
LocalMAP
1.10
1.12
1.09
1.09
1.11
1.08
1.30
1.11
1.13
Autoencoder
1.07
1.05
1.07
1.46
1.03
1.04
1.32
1.10
1.14
ParamPaCMAP
1.02
1.01
1.02
1.64
1.03
1.04
1.32
1.10
1.15
UMAP
1.17
1.15
1.11
1.19
1.13
1.13
1.38
1.14
1.17
Table 3: Order beyond membership: ω(10)=10⋅O-DLP(10)/DLP(10) , computed from Tables 4 and 5 . ω=1 is the order-blind baseline of Theorem 4 (ii), a neighborhood kept and shuffled; ω=10 is perfect order. At these sizes the baseline is exact to within 0.03 by Eq. ( 21 ) for every method with DLP≥0.02 , and to within 0.07 for all of them. The eleven peer methods that run at full size are scored on all eight benchmarks; NeRV is compared at matched scale in Table A.3 . Rows are sorted by the mean. Columns: MN MNIST, CF CIFAR-10, FM Fashion-MNIST, SC Speech Commands, RC Renal-Cortex, BC Brain-Cortex, CT Covertype, AG AG News.
Figure 3: Where the recovered order sits. Exact-hit rate at the first five positions: the fraction of points whose r -th input-space neighbor returns at position r in the embedding, on four benchmarks (log scale). This is the head of the list, where the position discount places its weight and where retrieval, label transfer and the k -nearest-neighbor graph read. SR-tSNE lifts every one of these positions on each of the four benchmarks shown. The equal-budget λ=0 control tracks t-SNE, so the lift is the rank term’s. All ten positions on all eight benchmarks are in Table A.2 .
Figure 4: End-to-end trading on IDCC at 2 bps (5-minute bars; identical DBSCAN + K hyperparameters). (a, c) Generated markers and (b, d) executed trades. The gold band highlights SR-tSNE’s two strongest round trips ( +12.5% vs. +2.6% for t-SNE); gold circles mark trade executions (dashed where t-SNE places no matching side). Over the quarter, SR-tSNE delivers +28.0% return at Sharpe 2.61 versus −7.5% at Sharpe −0.81 for t-SNE.
Figure 5: Cross-method comparison across 35 equities at 2 bps under identical DBSCAN + K scoring: (a) net return (%), (b) annualized Sharpe ratio, (c) maximum drawdown (%), and (d) round trips. Boxes indicate interquartile ranges; dots represent individual equities; top numbers denote medians; bottom numbers in (a) – (c) report two-sided paired Wilcoxon p -values versus SR-tSNE.
Figure 6: Retrieval comparison on (a) MNIST, (b) CIFAR-10, and (c) Fashion-MNIST across the top three embedded positions. Thumbnail badges indicate high-dimensional ground-truth ranks within the query’s top- 30 ; frame colors encode rank displacement (dark green: exact match; pale green: offset by one; orange: offset by ≥2 ; red: outside top- 30 ). Class labels turn red upon class mismatch. Queries illustrate exact-rank preservation where comparative methods introduce ordinal and semantic errors.
Dataset
t-SNE
PHATE
UMAP
TriMap
LocalMAP
DREAMS
SR-tSNE
ARI ↑
Trachea
.7841
.5082
.6771
.5988
.5982
.6001
.8221
Liver-MacParland
.5776
.5347
.5815
.5842
.5066
.5629
.6012
Kidney
.7179
.4276
.6688
.6874
.6855
.6943
.7499
Kidney-Park
.2069
.2247
.2143
.2214
.1792
.1685
.2961
Brain-Cortex
.1141
.1618
.1640
.1607
.1444
.1515
.2070
Table 6: Clustering performance across six annotated single-cell datasets under the Leiden upper-envelope protocol. Bold indicates the top method per cell. Sign counts and two-sided Wilcoxon signed-rank p -values evaluate SR-tSNE against each baseline across all datasets ( p<0.05 bolded).
Figure 7: Cluster count discrepancies relative to curated ground-truth annotations across six single-cell datasets. (a) Signed deviation per dataset (zero denotes the ground truth). (b) Mean absolute error decomposed into over-clustering (excess clusters, solid bars) and under-clustering (omitted clusters, hatched bars). SR-tSNE minimizes spurious cluster fragmentation.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A.1 : Cumulative return of the 35 names at 2 bps, one row per group. Buy and hold is charged the same round-trip cost.
Figure A.2 : What the markers are worth. The edge of an embedding on one name is the mean move of the close over the next 24 bars after its buy markers minus the mean move after its sell markers. a gives the two halves separately, each a median over the 35 names of Table A.12 ; b tracks the median edge at three holding horizons, with the paired Wilcoxon p against t-SNE beside each; c gives the per-name difference against t-SNE at two hours.
Figure A.3 : DLP versus neighborhood size k on the same eight benchmarks, with the legend at lower right. Each panel carries the methods for which a sweep across k was stored; the four methods added at k=10 in Table 5 are scored at that single size and are not swept here. SR-tSNE holds the top curve at k∈{5,10} on every benchmark.
Dataset
n
λ
UMAP
TriMap
t-SNE
LocalMAP
DREAMS
SR-tSNE
Δ
P 1 : is the first neighbor returned of the query’s class
MNIST
70 000
50
.8811
.8351
.9288
.9164
.9019
.9354
+0.0063
Fashion-MNIST
70 000
50
.7159
.6877
.8296
.7471
.8173
.8382
+0.0086
CIFAR-10
60 000
50
.7218
.6722
.8035
.7267
.7707
.8053
+0.0016
Tiny-ImageNet
70 000
50
.2479
.1689
.4561
.2706
.4044
.4584
+0.0023
EMNIST-Balanced
69 983
50
.4608
.3174
.6728
.5692
.6090
.6888
+0.0164
Appendix
Table A.20 : Label transfer across seven vision datasets, one paired observation each, with Δ taken against the equal-budget control ( λ=0 ), whose own score on each row is SR-tSNE minus Δ . Every query is scored from its ten nearest reference points in the two-dimensional embedding, on the same held-out fifth of each dataset. Bold marks the best score in a row. The two rows below each block give the sign count and the two-sided Wilcoxon signed-rank test over the seven datasets, where bold marks p<0.05 .